Jev × Foundry: Calibrated AI Decisions on the Ontology
Jev (TypeSafe AI, Sept 2026) is a System One model: it returns typed answers with calibrated probabilities instead of generated text. Learn how it works, how to design questions it answers well, how to earn confidence thresholds on your own data, and how to wire it into Palantir Foundry — Ontology in, Action out, with a Workshop review queue — ending with a technician-dispatch capstone. Written against the official jev-1.13.0 docs while Jev is in early access; assumes the Palantir and TypeScript for Foundry bootcamps. Bilingual slides and notes (English + 中文).
Premium content
Upgrade to premium to access all lessons, tutorials, and hands-on exercises.
Lesson Overview
What Is Jev? Decisions, Not Text
Meet Jev, TypeSafe AI's System One model: state and typed questions in, typed answers with calibrated probabilities out — no generated text. Learn the Choice / Score / Noul primitives, what calibration does and doesn't promise, Jev's documented weak spots, and why it fits Palantir Foundry's Ontology — with the two caveats that shape the rest of the series.
Your First Jev Call: State, Questions, Answers
Make your first Jev request: one maintenance work order as state, plus a Choice, a Score, and a Noul — sent with curl, the Python SDK (typesafe-sdk), and the TypeScript SDK (@typesafe-ai/sdk). Read every response field — choice, probabilities, confidence, score, legend, noul — and learn the error codes, batching, and why you pin jev-1.13.0.
Designing Questions Jev Answers Well
Nine rules for writing questions Jev answers well, drawn from TypeSafe's official guidance and applied to maintenance work orders: one snap judgment per question, choosing Choice vs Score vs Noul, contrastive criteria, keeping math and dates in code, structured instructions, decomposition with weights, speculative fan-out, why answers don't have to agree, and the language and adversarial-text risks.
Confidence, Calibration & Human Review
Turn Jev's probabilities into safe decisions: three confidence bands (act / confirm / human), thresholds that scale with risk, and gating in code. Then earn those thresholds — backtest on Foundry history with a reliability table, ECE, and Brier score, pick a gate from target accuracy vs coverage, and design a review queue whose overrides become labels.
Wiring Jev into Foundry: Ontology In, Action Out
Connect Jev to Palantir Foundry end to end: a new WorkOrder triggers Automate, a function-backed Action runs a TypeScript v2 Function that calls Jev through a Data Connection source, and gated edits come back where every status has an owner — emergencies page on-call, AUTO sets the trade, REVIEW goes to a coordinator, failures land in manual triage with the reason recorded. Plus the OSDK path for backfills and re-triage, governance and cost answers, and common pitfalls.
Capstone: Dispatch with Jev + Constraints
Answer the big question: can Jev decide which technician to dispatch — or which flight to book? Yes, as the judge, not the dispatcher: triage first, hard-filter candidates on the final trade in code, ask one fit Noul per candidate in a single request, score and assign in code (holding only the slots you take, with the Assign Action enforcing capacity), keep emergencies with on-call, and prove it in shadow mode against a baseline.